How AI impacts the workforce: An explanatory framework

Abstract

As artificial intelligence (AI) technologies advance rapidly, organisations face critical decisions about when to implement AI for workforce tasks and when to retain human workers. Existing research typically predicts task allocation towards AI or human labour by categorising jobs based on the types of activities they involve. However, these predictions frequently fail to explain why similar tasks are automated in some organisations but remain manual in others, creating a significant gap between existing research and real-world implementation decisions.

This thesis addresses this gap by investigating explanations for how organisations actually make AI implementation decisions in practice. Rather than starting with predetermined task categories, the study uses semi-structured interviews across healthcare, legal services, logistics, manufacturing, and government sectors to examine examples of where AI has been successfully deployed and where it remains absent. This qualitative approach, underpinned by Critical Realism, seeks to understand the underlying factors that explain these implementation patterns.

The analysis reveals that task allocation between AI and human workers depends on three key elements which are captured within an explanatory framework. Economic dynamics create organisational pressure to seek technological solutions to business challenges. Operational leverage points identify where human limitations create bottlenecks that AI might address. Most importantly, operational execution complexity determines whether AI can actually perform reliably in specific contexts, based on factors such as data availability, task uncertainty, and acceptable risk levels.

A critical finding is that the context in which the work is performed, rather than simply what the work involves, fundamentally shapes whether AI implementation succeeds. Tasks that appear similar when described by their activities may have a very different automation potential depending on their situational context, including the level of uncertainty, the judgement required, and the availability of data. This explains why automation predictions based solely on task descriptions often prove inaccurate.

These insights are consolidated into a framework that organisations can use to assess where AI implementation is likely to succeed. This framework recognises that implementation success depends on the interaction between organisational needs, operational constraints, and the situational context of the task, while establishing foundations for developing more detailed substantive theories that identify causal mechanisms. This research contributes theoretically to understanding AI-workforce relationships and provides practical guidance for navigating AI implementation decisions.

Date of Award	10 Sept 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Hugh Lauder (Supervisor) & Linda Newnes (Supervisor)